gather_nd_op.cc 5.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include "paddle/fluid/operators/gather_nd_op.h"
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/ddim.h"

namespace paddle {
namespace operators {

class GatherNdOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE_EQ(ctx->HasInput("X"), true,
                      "Input(X) of GatherNdOp should not be null.");
    PADDLE_ENFORCE_EQ(ctx->HasInput("Index"), true,
                      "Input(Index) of GatherNdOp should not be null.");
    PADDLE_ENFORCE_EQ(ctx->HasOutput("Out"), true,
                      "Output(Out) of GatherNdOp should not be null.");

    auto x_dims = ctx->GetInputDim("X");
    auto x_dims_size = x_dims.size();
    auto index_dims = ctx->GetInputDim("Index");
    auto index_dims_size = index_dims.size();

41 42 43
    PADDLE_ENFORCE_LE(
        index_dims[index_dims_size - 1], x_dims_size,
        "Input(Index).shape[-1] should be no greater than Input(X).rank");
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183
    PADDLE_ENFORCE_GE(index_dims_size, 2UL,
                      "The rank of Input(Index) should be greater than 1");

    std::vector<int64_t> result_dims;
    // The result dims is
    //   Index.shape[:-1] + X.shape[Index.shape[-1]:]
    for (int i = 0; i < index_dims_size - 1; ++i) {
      result_dims.emplace_back(index_dims[i]);
    }
    for (int i = index_dims[index_dims_size - 1]; i < x_dims_size; ++i) {
      result_dims.emplace_back(x_dims[i]);
    }

    ctx->SetOutputDim("Out", framework::make_ddim(result_dims));
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
                                   ctx.device_context());
  }
};

class GatherNdGradOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
    ctx->ShareLoD("X", /*-->*/ framework::GradVarName("X"));
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        ctx.Input<Tensor>(framework::GradVarName("Out"))->type(),
        ctx.device_context());
  }
};

class GatherNdOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
    AddInput("X", "The source input of gather_nd op");
    AddInput("Index", "The index input of gather_nd op");
    AddOutput("Out", "The output of gather_nd op");
    AddComment(R"DOC(
    Gather_Nd Operator.

    This function is actually a high-dimensional extension of gather 
    and supports for simultaneous indexing by multiple axes. Out is 
    obtained by gathering slices from X into a tensor with shape 
    Index.shape[:-1] + X.shape[Index.shape[-1]:].

    Example:
   
    Given:
         X = [[[ 0,  1,  2,  3],
               [ 4,  5,  6,  7],
               [ 8,  9, 10, 11]],
              [[12, 13, 14, 15],
               [16, 17, 18, 19],
               [20, 21, 22, 23]]]
       
         X.shape = (2, 3, 4)

   *Case 1:

       Index = [[1]]

    we get:
       Out = 
            [[12, 13, 14, 15],
             [16, 17, 18, 19],
             [20, 21, 22, 23]]

   *Case 2:

       Index = [[0,2]]

    we get:
        
       Out =  [8, 9, 10, 11]

   *Case 3:

       Index = [[1, 2, 3]]

    we get:

       Out = [23]

)DOC");
  }
};

class GatherNdGradOpDescMaker : public framework::SingleGradOpDescMaker {
 public:
  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;

 protected:
  std::unique_ptr<framework::OpDesc> Apply() const override {
    std::unique_ptr<framework::OpDesc> op(new framework::OpDesc());
    op->SetType("gather_nd_grad");
    op->SetInput("Index", Input("Index"));
    op->SetInput("X", Input("X"));
    op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
    op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
    op->SetAttrMap(Attrs());
    return op;
  }
};

DECLARE_NO_NEED_BUFFER_VARS_INFERENCE(GatherNdGradNoNeedBufferVarInference,
                                      "X");

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;

REGISTER_OPERATOR(gather_nd, ops::GatherNdOp, ops::GatherNdOpMaker,
                  ops::GatherNdGradOpDescMaker);

REGISTER_OPERATOR(gather_nd_grad, ops::GatherNdGradOp,
                  ops::GatherNdGradNoNeedBufferVarInference);

REGISTER_OP_CPU_KERNEL(gather_nd, ops::GatherNdOpKernel<float>,
                       ops::GatherNdOpKernel<double>,
                       ops::GatherNdOpKernel<int64_t>,
                       ops::GatherNdOpKernel<int>,
                       ops::GatherNdOpKernel<uint8_t>);

REGISTER_OP_CPU_KERNEL(gather_nd_grad, ops::GatherNdGradOpKernel<float>,
                       ops::GatherNdGradOpKernel<double>,
                       ops::GatherNdGradOpKernel<int64_t>,
                       ops::GatherNdGradOpKernel<int>,
                       ops::GatherNdGradOpKernel<uint8_t>);